Rough Isometries of Order Lattices and Groups
Dissertation zur Erlangung des mathematisch-naturwissenschaftlichen Doktorgrades “Doctor rerum naturalium” der Georg-August-Universit¨at G¨ottingen
vorgelegt von Andreas Lochmann aus Bad Pyrmont
G¨ottingen 2009 Referenten der Dissertation: Prof. Dr. Thomas Schick Prof. Laurent Bartholdi, Ph.D.
Mitglieder der Pr¨ufungskommission: Prof. Dr. Thomas Schick Prof. Laurent Bartholdi, Ph.D. Prof. Dr. Rainer Kress Prof. Dr. Ralf Meyer Prof. Dr. Ulrich Stuhler Prof. Dr. Andreas Thom
Tag der m¨undlichen Pr¨ufung: 6. August 2009 Gewidmet meinen Eltern
Preface
Assume you measure the distances between four points and find the following: One of the distances is 143, the distance between the other pair of points is 141, and all other distances are 99. You know that your error in measurement is not greater than 1. Is this result then compatible with the hypothesis, that the four points constitute a square in Euclidean space? Another example: You distribute five posts equidistantly along straight rails, then watch a train driving on the rails and take the times at which it passes the posts. The times you measure are: 0, 100, 202, 301, 401. However, the time the clocks show might deviate by up to 1 from the clock at the starting post. Are the measured times consistent with the assumption that the train has constant speed in Newtonian mechanics? These two puzzles are metaphors for a general problem in the application of mathematics. We have to use measurements that might contain errors to verify predictions of an ideal theory. If, for example, you find a function f between vector spaces to fulfill linearity only approximately, i.e. there is ǫ > 0 such that
d f(x + y), f(x) + f(y) ǫ ≤ for all x, y; can you find a perfectly linear function g which is near f? Ulam stated this question during a talk around 1941, and Hyers found a positive solution shortly after by applying scaling arguments ([Hy]). Ulam and Hyers continued their collaboration, and were able to give positive answer in [HU] to the following question in 1947: Is every ǫ-isometric homomorphism of real- valued continuous function spaces in bounded distance to an isometry?
Theorem 1 (Hyers-Ulam) 1 Let K, K′ be compact metric spaces and let E and E′ be the spaces of all real valued continuous functions on K and K′, respectively. Let T (f) be a homo- morphism of E onto E′ which is also an ǫ-isometry for some arbitrary, fixed ǫ 0. Then there exists an isometric transformation U(f) of E onto E such ≥ ′ that U(f) T (f) 21ǫ for all f in E. In particular, the underlying metric || − || ≤ spaces K and K′ are homeomorphic.
This theorem evoked several similar results for linear functions ([Ra2]), but also about stability of differential equations, the stability of group actions ([GKM]), and approximate group homomorphisms, as defined by Ulam in [U], section VI.1; see [HR] for a survey on this topic. 6
It is interesting to see that the isometry of the function spaces implies con- ditions of their underlying metric spaces—in this case they are homeomorphic. Our main interest during a large part of this thesis will be function spaces of Lipschitz functions, and their stability against rough isometries. Rough isometries are in some sense the prototype of approximative struc- ture: the description of Hyers’ ǫ-linear functions would not be possible without a metric on the vector spaces. The idea simply is to replace an identity x = y by d(x, y) ǫ. In this sense, each description of approximations needs metric ≤ spaces. As the natural mappings of metric spaces are Lipschitz maps, isome- tries, and isometric embeddings, their rough counterparts are the most natural approximative tools, among them the rough isometry (or ǫ-isometry) as mea- sure to describe the similarity of metric spaces. Our goal is to shed some light on the interplay of rough isometries with order lattice and group structures, but in particular to demonstrate their use in the study of the geometry of Lipschitz function spaces. There are detailed books and various articles on many aspects of Lipschitz function spaces, including isometries between them. A nice survey is Weaver’s book on Lipschitz Algebras [Wv] (cf. Section 2.6, where Weaver elaborates on the exact same questions we want to tackle here, but in a non-coarse context and under different conditions). However, to the knowledge of the author, no book nor article dealt with their coarse geometry yet. On the other hand, Lipschitz functions naturally appear in many aspects of coarse geometry, such as the Levy concentration phenomenon or the definition of Lipschitz-Hausdorff distance in [Gv2]. But they are not dealt with as metric spaces either. The structure of this thesis is as follows: We first give an overview of basic notions in handling with metric spaces and order lattices. We assume that the reader is familiar with metric spaces, so this part will just cover our use of infinities in metric spaces, the definition of hyperconvex and injective metric spaces, and those notions from coarse geometry which apply to our situation. The first chapter deals with a generalization of Birkhoff’s metric lattices to put the supremum metric on Lipschitz function spaces into a common context. During this chapter we derive several useful tools for to handle distances in Lipschitz function spaces, and introduce several metrics for Lipschitz functions, familiar ones as well as some more exotic distance functions. The final section in this chapter will define a very important notion for our analysis, a version of irreducibility which depends on the chosen metric of a lattice. We further see some first, general properties of this new kind of irreducibility. The second chapter will concentrate on Lipschitz function spaces with supre- mum metric, and we will examine the question, whether they are roughly iso- metric when their underlying spaces are. We first take a look at a simple, but efficient smoothening algorithm, then take a digression to demonstrate a corresponding Lemma for hyperconvex-space-valued Lipschitz functions. We then revive the situation of the first chapter, and define a rough ver- sion of isomorphism for intervaluation metric lattices, which fits well into the context of our new irreducibility-condition. By introducing minimal Lipschitz functions (which we call “Λ-functions”), we provide a lattice-theoretic base of the Lipschitz function space (see Definition 76), and put the smoothening re- Preface 7 sult of the first section into this more advanced context, re-proving it using new techniques:
Theorem 2 2 Let X,Y be metric spaces, ǫ 0. For each ǫ-isometry η : X Y , there is a + ≥ → 4ǫ-ml-isomorphism κ : Lip Y Lip X such that κ is ǫ-near f f η for all → 7→ ◦ f Lip Y (see Definitions 6, 29, 80, 89). ∈ We then continue to analyze our Λ-functions and find them to be exactly the irreducible elements of the first chapter. This allows us to state a converse of Theorem 2:
Theorem 3 3 Let X,Y be complete metric spaces and ǫ 0. For each ǫ-ml-isomorphism + ≥ κ : Lip Y Lip X there is an 88ǫ-isometry η : X Y , such that κ is 62ǫ-near → → f f η for all f Lip Y . 7→ ◦ ∈ As we make use of the order-lattice structure of the Lipschitz function spaces in our proofs, the following theorem by Kaplansky [Kp] is related to our results as well and we state it in its formulation by Birkhoff ([Bi1], 2nd ed. p. 175f.):
Theorem 4 (Kaplansky) 4 Any compact Hausdorff space is (up to homeomorphism) determined by the lattice of its continuous functions.
However, the proofs we present here not only make use of the lattice struc- ture of Lip X, but of a metric on it as well. In this sense, the comparison with Kaplansky’s Theorem 4 is not perfect. We then ask whether it is possible to state Theorems 2 and 3 in a more general context, and give counter-examples to a broad range of different metrics on Lipschitz function spaces, among them the seemingly convenient L1-metric. We close this chapter with an example concerning quasi-isometries, and an application in the theory of scaling limits. The third chapter discusses the use of rough isometries in the context of finitely generated groups. One cannot overvalue the importance of coarse meth- ods in Geometric Group Theory, which is in large parts based on the notion of quasi-isometry. We begin this chapter with another digression, demonstrating how some proofs can be reformulated in a rough context, while this is not possi- ble with others. We then try to point out the possibilities rough isometries add to Geometric Group Theory, by providing a specifically rough isometry invariant (the exponential growth rate), and then show the connection between symme- tries in some virtually abelian groups and rough isometries between them. We then continue with a theorem motivated by Bartholdi and translate it into the rough context in the final section, where we show that uniformly rough and quasi-isometries imply commensurability. Finally a remark to our notation: N0 denotes the non-negative integers, N∗ the positive integers, Cn denotes the cyclic group of order n. We sometimes drop function brackets where feasible. 8
Acknowledgements. The author wants to thank the “Graduiertenkolleg Gruppen und Geometrie” for supporting his research financially, through con- ferences and workshops, and by supporting the great open atmosphere at the institute. Best thanks go to Prof. Dr. Thomas Schick, Prof. Laurent Bartholdi, and Prof. Dr. Andreas Thom for their friendly support and great expertise, and to Prof. Dr. Themistocles Rassias for pointing the author at important works in connected fields. The author thanks his family and friends for al- ways supporting him through the course of the doctorate, and in particular Dr. Johannes H¨artel for being friend, critic, and guide.
Danksagung. Der Autor m¨ochte dem “Graduiertenkolleg Gruppen und Geometrie” daf¨ur danken, dass es seine Forschung sowohl finanziell als auch durch Konferenzen und Workshops und durch den Erhalt der großartigen At- mosph¨are am Institut unterst¨utzt hat. Bester Dank geht an Prof. Dr. Thomas Schick, Prof. Laurent Bartholdi, Ph.D., und Prof. Dr. Andreas Thom f¨ur ihre freundliche Unterst¨utzung und große Expertise, sowie an Prof. Dr. Themistocles Rassias f¨ur die Verweise auf wichtige Arbeiten in verwandten Gebieten. Der Autor dankt seiner Familie und seinen Freunden f¨ur die stete Unterst¨utzung im Laufe des Doktorats, und insbesondere Dr. Johannes H¨artel f¨ur Freundschaft, Kritik und Anleitung.
Georg-August-Universit¨at G¨ottingen, Germany, 2009 eMail [email protected] Contents
0 Basic Notions 11 0.1 Basic Notions in Metric Spaces ...... 11 0.1.1 InfiniteMetrics ...... 11 0.1.2 Injective and Hyperconvex Metric Spaces ...... 13 0.1.3 Basic Notions of Coarse Geometry ...... 16 0.2 Basic Notions in Order Lattices ...... 17
1 Order Lattices with Metrics 23 1.1 Valuation Metric Lattices ...... 23 1.2 Ultravaluation Metrics ...... 32 1.3 Intervaluation Metrics ...... 34 1.4 Complete Metric/Lattice-Irreducibility ...... 39
2 Rough Isometries of Lipschitz Function Spaces 43 2.1 SmootheningofLipschitzFunctions ...... 43 2.1.1 Hyperconvex and Non-Hyperconvex Codomains ...... 48 2.1.2 Algorithmic Aspects ...... 51 2.2 Roughml-Isomorphisms ...... 51 2.3 Λ-Functions...... 54 2.4 Λ-Functions are Completely ml-Irreducible ...... 59 2.5 InducingRoughIsometries...... 61 2.6 Other Metrics for Lip X ...... 65 2.7 Quasi-Isometries ...... 67 2.8 Scalinglimits ...... 68
3 Rough Isometries of Groups 71 3.1 TheTheoremofMazur-Ulam ...... 71 3.1.1 RoughAbelianness...... 75 3.2 CoarseRelationsforGroups...... 75 3.3 ExponentialGrowthRate ...... 77 3.4 TheAbelianCase...... 78 3.5 Rough Isometries of Quotients with Finite Kernel ...... 81 3.6 SharedIsometries...... 82 3.7 SharedRoughandQuasi-Isometries ...... 87
Bibliography 95 10 Chapter 0
Basic Notions
0.1 Basic Notions in Metric Spaces
0.1.1 Infinite Metrics
As we will deal with lattice-theoretic notions, it is convenient to ensure the existence of a greatest possible distance in a metric space. To do so without having to restrict to bounded metric spaces, we will extend our notion of metric spaces to include infinite distances.
Definition 5 ...... 5 We define the binary operation + on [0, ] to be as expected on [0, ), and ∞ ∞ set + z = z + = as well as z for all z [0, ]. We call ∞ ∞ ∞ ≤ ∞ ∈ ∞ z [0, ] positive if z = 0. ∈ ∞ 6
Definition 6 ...... 6 A pseudo-metric+ space (X, d) is a non-empty set X together with a mapping d : X X [0, ] which has the following properties: × → ∞
1. d(x, x) = 0 for all x X, and ∈ 2. d(x, z) d(x, y) + d(y, z) for all x, yz X. ≤ ∈
A pseudo-metric+ space (X, d) is a metric+ space if it is positive-definite, i.e.
3. d(x, y) = 0 implies x = y for all x, y X. ∈
The mapping d is then called a “metric”, “metric+ ”, or “distance”.
A (pseudo-)metric+ space (X, d) is a (pseudo-)ultrametric+ space, if it fulfills
1. d(x, z) d(x, y) d(y, z) for all x, yz X, ≤ ∨ ∈ 12 Claim 8 where denotes the maximum. ∨ A (pseudo-)metric+ space (X, d) is called a true (pseudo-)metric space if d(x, y) = for all x, y X. This equals the common notion of a metric space in the 6 ∞ ∈ literature, like [BBI].
Definition 7 ...... 7 The diameter diam A of a non-empty subset A in a pseudo-metric+ space is the supremum of all distances of pairs of points in A.
A bounded (pseudo-)metric space is a (pseudo-)metric+ space X with finite diameter diam X < . ∞
Definition 8 ...... 8 An isometry [isometric embedding] is a bijection [injection] between pseudo-me- tric+ spaces which preserves the distance.
In most cases we call metrics just “ d ” without reference to the metric space, as it should be clear from the elements which metric is meant. We will make heavy use of the symbol
d(x,x′) d(y,y′) z − ≤
for some x, x′, y,y′ in X or Y and z [0, ]. To make sense of this in case one ∈ ∞ of the distances becomes infinite, we define the former symbol to be equivalent to
d(x, x′) d(y, y′) + z and d(y, y′) d(x, x′) + z. ≤ ≤ In particular, we find = 0. This might seem unfamiliar. Note however, |∞ − ∞| that z z can be perfectly understood as a metric on [0, ] itself. | − ′| + ∞ Remark There is no non-trivial convergence to in this metric, is just an ∞ ∞ infinitely far away point. We will not need a convergence to infinity anyway; the addition of serves to obtain completeness of order lattices, particularly ∞ of [0, ] itself. ∞ Furthermore, it is obvious that a metric+ space X always is a disjoint union of true metric spaces X with d(x, x ) = if and only if x X , x X with j ′ ∞ ∈ j ′ ∈ k j = k, for j, k J. We call the X components of X. We call X complete, 6 ∈ j if all of its components are complete as true metric spaces, i.e. each Cauchy sequence converges.
The difference between metric+ and true metric spaces is rather small, as a metric+ space merely is a collection of true metric spaces. For example, the topology of a metric+ space is the disjoint union of the topology of its com- ponents, and any converging sequence eventually is contained in one of these components, with only finitely many exceptions at the beginning of the se- quence. Basic Notions in Metric Spaces 13
Example 9 ...... 9 The algebra of continuous functions f : R R together with the supremum → metric
d (x, y) := sup f(x) f(y) ∞ x R − ∈ is a metric+ space. Its components are those functions with finite distance to each other, so the subset of bounded functions constitutes the component of the zero function, and the identity belongs to another component.
Example 10 ...... 10 Similar to the usual definition of a norm on a vector space, it is possible to define a norm+ with possibly infinite value, and derive a metric+ from it. For example, the true metric space of square-summable complex sequences is a component of the metric+ space of all complex sequences with norm+
(x , x ,...) := x 2 [0, ] 1 2 2 | j| ∈ ∞ sj N∗ X∈ for any complex sequence (x1, x2,...).
Definition 11 ...... 11 The Hausdorff-distance of two subsets A and B in a pseudo-metric+ space is the infimum of all non-negative numbers r such that for each x A there is ∈ y B with d(x, y) r and vice versa. ∈ ≤ A subset A of a pseudo-metric+ space is ǫ-dense in A, if its Hausdorff-distance to A is less than or equals ǫ.A dense subset is a 0-dense subset. A roughly dense subset is a subset for which its Hausdorff-distance to A is finite.
Definition 12 ...... 12 A pseudo-metric+ space X is separable if it contains a countable dense subset.
0.1.2 Injective and Hyperconvex Metric Spaces
Definition 13 ...... 13 A short map (sometimes called a “metric map”) between pseudo-metric+ spaces X and Y is a 1-Lipschitz map π : X Y , this means it fulfills d (πx, πy) → Y ≤ d (x, y) for all x, y X. X ∈ Short maps are the canonical morphisms to construct a category Met of true metric spaces, as a bijective short map whose inverse is short as well is an isometry. They form morphisms in the category Met+ of metric+ spaces as well, as short maps map componentwise.
Definition 14 ...... 14 An injective metric+ space X is an injective object in the category Met+, i.e. 14 Claim 18 for each metric space Y and short map f : Y X, and each short embedding + → i : Y ֒ Z of Y into another metric space Z, there is a short map g : Z X → + → which extends f (f = g i). ◦ Proposition 15 ...... 15 A metric+ space is injective if and only if all of its components are injective. In particular, a true metric space is injective in Met if and only if it is injective in Met+.
Proof Note that a short map maps a component into a single component of the codomain, i.e. the map which assigns each metric+ space its set of components is a functor. “ ” Let X be a component of the injective metric space X. Let f : Y X ⇒ 1 + → 1 be a short map, and Y Z metric spaces. Y must be a true metric space. ⊆ + Let Z be the component of Y in Z. Extend f to the whole of X, its image 1 |Z1 must still be in the same component X . Now choose an arbitrary point x X 1 0 ∈ 1 and map all remaining components of Z to this point. This yields an extension g : Z X of f. → 1 “ ” Let Y Z, f : Y X, where X is componentwise injective. Choose a ⇐ ⊆ → point x X. Extend f componentwise on each component of Z which inter- 0 ∈ sects Y ; map each remaining component of Z to x0.
As next, we state the classical result that metric spaces are injective if and only if they are hyperconvex. This accounts for true metric spaces as well as for metric+ spaces. Hyperconvexity is a stronger form of convexity.
Definition 16 ...... 16 Let X be a metric+ space. X is convex in the sense of Menger ([Me], p. 81) if for each x,y X there is z X r x, y such that d(x, y)= d(x, z)+ d(z, y). ∈ ∈ { } X is totally convex (or convex in the sense of [EK]) if for all a, b (0, ] with ∈ ∞ d(x, y) a + b there is z X with d(x, z) a and d(z, y) b. ≤ ∈ ≤ ≤ X is hyperconvex if for any family of points (xj)j J X and numbers rj [0, ] ∈ ⊆ ∈ ∞ with d(x , x ) r + r for all j J, with J an arbitrary index set, there is j k ≤ j k ∈ an element z X with d(x , z) r for all j J. In simpler words: If any ∈ j ≤ j ∈ family of balls could theoretically intersect pairwise (given their radii), they all intersect.
Example 17 ...... 17 Typical examples for hyperconvex true metric spaces are (real) trees and Rn with L∞-metric as well as certain subsets thereof. In particular, R and each real interval are hyperconvex.
Proposition 18 ...... 18 A metric+ space X is hyperconvex if and only if all of its components are hy- perconvex. Basic Notions in Metric Spaces 15
Proof All xj with infinite rj can be neglected. On the other hand, if two points x ,x X are in different components, at least of one r or r must be j k ∈ j k infinite, hence the hyperconvexity property always reduces to hold on a single component of X. On the other hand, assume X is hyperconvex. Given a family of balls with center x and radius r in one component X X, we may use j j 1 ⊆ hyperconvexity in X to gain the desired element z X. As long as at least two ∈ of the r were finite, we must choose z X . If only one r is finite, we may j ∈ 1 j choose z = xj.
To provide a feeling for hyperconvex and injective metric+ spaces, we cite some assorted theorems without giving proofs. Due to Propositions 15 and 18, there are no obstructions to generalize the original statements to metric+ spaces.
Theorem 19 ([AP], Theorem 2.4) 19 A metric+ space is hyperconvex if and only if it is injective.
Theorem 20 ([EK], Theorem 3.1) 20 A product of arbitrarily many hyperconvex metric+ spaces with supremum me- tric+ is hyperconvex. (This statement is much easier in the Met+ form than the original statement for true metric spaces.)
Theorem 21 ([EK], Proposition 3.2) 21 Any hyperconvex metric+ space is complete.
Theorem 22 (Baillon’s Theorem; [EK], Theorem 5.1) 22 The intersection of a descending chain of non-empty hyperconvex subsets of a bounded metric space is non-empty and hyperconvex.
Theorem 23 ([EK], Theorem 6.1) 23 The fixed point set of a short map f : X X acting on a hyperconvex bounded → metric space X is non-empty and hyperconvex.
Theorem 24 ([EK], Theorem 9.6) 24 Let X be hyperconvex, K > 0. The family of all bounded K-Lipschitz functions X X with supremum metric is hyperconvex. → Theorem 25 ([I], Theorem 2.1) 25 Each true metric space X has an injective envelope (i.e. a minimal injective space into which X isometrically embeds), which is unique up to isometry.
Theorem 26 ([I], Remark 2.11) 26 The injective envelope of a compact space is compact.
Theorem 27 ([AP], Theorem 3.3) 27 If X is a metric space, A X hyperconvex, B X such that A and B have + ⊆ ⊆ identical closures in X, then B is hyperconvex as well. 16 Claim 31
Theorem 28 ([AP], Theorem 3.9) 28 Let X be hyperconvex. Then A X is hyperconvex if and only if it is a retract ⊆ of X by a contracting retraction.
0.1.3 Basic Notions of Coarse Geometry
A well written introduction to coarse geometry is the book by Burago, Burago and Ivanov ([BBI]). In the following, let X and Y be metric+ spaces.
Definition 29 ...... 29 Two (set theoretic) mappings α, β : X Y are ǫ-near to each other, ǫ 0, if → ≥ d (αx,βx) ǫ x X. (We drop brackets where feasible.) Y ≤ ∀ ∈ A (set theoretic) mapping α : X Y is ǫ-surjective, ǫ 0, if for each y Y → ≥ ∈ there is x X such that d (αx,y) ǫ. ∈ Y ≤
Definition 30 ...... 30 A (not neccessarily continuous) map η : X Y is called a (λ, ǫ)-quasi- → isometric embedding, ǫ, λ 0 (which shall always imply λ, ǫ R), if ≥ ∈ 1 λ− d (x, x′) ǫ d (ηx, ηx′) λ d (x, x′) + ǫ X − ≤ Y ≤ X for all x,x X. ′ ∈ A pair η : X Y , η : Y X of (λ, ǫ)-quasi-isometric embeddings is called a → ′ → (λ, ǫ)-quasi-isometry if η η and η η are ǫ-near the identities on Y and X, ◦ ′ ′ ◦ respectively. When we speak of a “quasi-isometry η : X Y ” a corresponding → map η′ shall always be implied. X and Y are called quasi-isometric, if there is a quasi-isometry between them.
Definition 31 ...... 31 A (1, ǫ)-quasi-isometric embedding, ǫ 0, is called a ǫ-isometric embedding. It ≥ fulfills
d (x,x′) d (ηx,ηx′) ǫ | X − Y | ≤ for all x,x X. ′ ∈ An ǫ-isometry is a (1, ǫ)-quasi-isometry. The map η : X Y is called a rough → isometry if there is some ǫ 0 such that η is an ǫ-isometry. ≥ X and Y are called ǫ-isometric [roughly isometric], if there is an ǫ-isometry [any ǫ-isometry] between them.
Historical Remark It is difficult to attribute the concept of rough isometry to a single person, as it was always present in the notion of quasi-isometry, which itself was an obvious generalization of what was then called pseudo-isometry by Mostow in his 1973-paper about rigidity (see [Mo], [Gv1], [Kn]). Recent developments about the stability of rough isometries can be found in [Ra1]. Basic Notions in Order Lattices 17
Definition 32 ...... 32 A mapping f : X Y is called (K, ǫ)-Lipschitz (i.e. “K-Lipschitz map on → ǫ-scale” in [Gv3]), ǫ, K 0, iff ≥ d (f(x), f(y)) K d (x, y) + ǫ x, y X. Y ≤ · X ∀ ∈
If ǫ = 0, f is K-Lipschitz (continuous). Define LipK,ǫ(X, Y ) to be the set of all (K, ǫ)-Lipschitz functions X Y , and Lip X := Lip (X, [0, ]), → K,ǫ K,ǫ ∞ Lip X := Lip1,0(X).
If nothing else is said, [0, ] = R 0 is the default codomain for a ∞ ≥ ∪ {∞} Lipschitz function. Assume f to be a (K, ǫ)-Lipschitz function on X and f(x) = for some ∞ x X. Then clearly f(y) = for all y in finite distance to x. Thus, if X is a ∈ ∞ true metric space, we have Lip X = Lip(X, R 0) . ≥ ∪ {∞}
0.2 Basic Notions in Order Lattices
Good starting points for Lattice Theory are Gr¨atzer ([Gr]) and the classical book by Birkhoff ([Bi1]).
Definition 33 ...... 33 A lattice (L, , ) is a set L together with two mappings , : L L L ∧ ∨ ∧ ∨ × → which are commutative, associative and fulfill the absorption laws f (f g)= ∧ ∨ f (f g)= f for all f, g L. A lattice is a partially ordered set by ∨ ∧ ∈
by absorption f g : f g = f f g = g ≤ ⇔ ∧ ⇔ ∨
L is distributive if and distribute over each other. ∨ ∧ L is bounded (as a lattice) if there exists a smallest element 0 L and a largest ∈ element 1 L. ∈ L is complete if all infima and all suprema of all subsets of L exist in L. (A complete lattice always is bounded.) L is complemented if L is bounded and for each f L there is a complement ∈ g L such that f g = 1 and f g = 0. ∈ ∨ ∧ L is a Boolean lattice if it is distributive and complemented.
Example 34 ...... 34 Let X be some set, and let L be any family of subsets of X closed under and ∩ . Then (L, , ) is a distributive lattice, sometimes called a “ring of sets” (we ∩ ∩ ∪ will stick to “lattice of sets”). If for all A L the complement X r A L as ∈ ∈ well, then (L, , ) is a Boolean lattice (“field of sets”). ∩ ∪ 18 Claim 38
Example 35 ...... 35 Let X be a topological space, and T the family of all open sets in X. T is a lattice by union and intersection, bounded by and X, distributive (as it is a ∅ lattice of sets). However, if X is Hausdorff, then T is complete if and only if T is the discrete topology, and if and only if T is complemented.
Definition 36 ...... 36 Let (L, , ) be a lattice. ∧ ∨ An element p L is join-irreducible if, whenever p = f g with f, g L, then ∈ ∨ ∈ p = f or p = g. An element p L is join-prime if, whenever p f g with f, g L, then p f ∈ ≤ ∨ ∈ ≤ or p g. ≤ An element p L is completely join-irreducible if, whenever p = f , with ∈ j J j f L, then p = f for some j J, J an arbitrary index set. Same∈ for com- j ∈ j ∈ W pletely join-prime. A lower set in a partially ordered set P is a subset Q of P with: f g, g Q, ≤ ∈ f P implies f Q. ∈ ∈ A sublattice of L is a subset of L closed under and . ∧ ∨ An ideal is a sublattice I L such that f g I whenever f L and g I ⊆ ∧ ∈ ∈ ∈ (equivalent definition: a sublattice which is a lower set). A proper ideal is an ideal I which is a proper subset of L. A principal ideal is an ideal I which is generated by a single element. A prime ideal is a proper ideal P of a Boolean algebra for which holds: If f, g L ∈ and f g P , then f P or g P . The family of all prime ideals in L is called ∧ ∈ ∈ ∈ P (L).
Proposition 37 ...... 37 In a distributive lattice, join-irreducibility and join-primeness are equivalent ([Bi1], Lemma III.3.1).
Proof Let p L be join-prime. Then it is join-irreducible by definition. Now ∈ let p L be join-irreducible and p f g with f, g L. We find ∈ ≤ ∨ ∈ p = p (f g) = (p f) (p g). ∧ ∨ ∧ ∨ ∧ As p is join-irreducible, we have p = p f, this is p f, or p = p g, which ∧ ≤ ∧ means p g. ≤
Example 38 ...... 38 We note that a prime ideal is not the same as a principal ideal generated by a join-prime element: The positive integers equipped with least common divisor and greatest common multiple constitute a distributive and unbounded lattice L = (N∗, gcd, lcm) Basic Notions in Order Lattices 19 with 1 as least element. Join-irreducible/join-prime elements are exactly the prime powers. The subset A = 5, 15, 50, 150 is an example for a sublattice { } in L. The least ideal encompassing A is the subset of all divisors of 150. It is a principal ideal generated by 150. The principal ideal which is generated by a join-irreducible element pn, p prime, is just 1, p, p2,..., pn . The subset { } B of all powers of 7 is a non-principal proper ideal. It is not a prime ideal: gcd(14, 35) = 7, but neither 14 B nor 35 B. The subset C = Zr 7Z of ∈ ∈ all positive integers except the multiples of 7 is a prime ideal: If 7 ∤ gcd(f, g), then 7 ∤ f or 7 ∤ g. On the other hand, the subset Zr 6Z is not even an ideal.
Example 39 ...... 39 Each ideal in a lattice is a lower set, but not vice versa: In the lattice L = (N , gcd, lcm) the subset Q := 1, 5, 25, 29 is a lower set, but not an ideal, ∗ { } because lcm(5, 29) = 145 / Q. ∈
Example 40 ...... 40 Let X be any set of at least two elements, and consider its powerset (L = ℘(X), , ). Its join-irreducibles are the one-element subsets and the empty ∩ ∪ set. A principal ideal generated by A X is ℘(A) ℘(X). Given any x X ⊆ ⊆ ∈ consider := A X : x / A . is a prime ideal. Fx { ⊆ ∈ } Fx
In Lipschitz function spaces, principal ideals which are generated by a join- irreducible are more interesting, as join-irreducibles tend to interact in more interesting ways: Their intersections can be non-trivial.
Theorem 41 41 (Representation Theorem for Distributive Lattices, G. Birkhoff 1933, M. H. Stone 1936; [Gr], Th. II.1.19) A lattice is distributive if and only if it is isomorphic to a lattice of sets.
Proof Birkhoff proved a very similar representation theorem for finite dis- tributive lattices in [Bi1] (see next Theorem), Stone later extended the proof to infinite distributive lattices as well. In addition, he showed that a comple- mented distributive lattice is isomorphic to a complemented lattice of sets. The main idea of the proof is to show that the map
π : L ℘(P (L)) → f p P (L) : f / p 7→ { ∈ ∈ } between L and the power set of the set P (L) of all prime ideals in L is an injective homomorphism, and hence π constitutes an isomorphism between |im π L and a sublattice of ℘(P (L)).
Subsequently, each distributive lattice is isomorphic to a sublattice of a power- set. Today there is a broad variety of representation theorems for (distributive) lattices, particularly as lattices of open, closed or clopen subsets in topological 20 Claim 45 spaces. These theorems are subsumed under the term “Stone-type dualities”. An earlier version is the following theorem by Birkhoff for finite distributive lattices. We will not make use of it, but its main idea of reconstructing a distributive lattice solely from its join-irreducibles will be a major theme in Chapter 2.
Theorem 42 (Birkhoff’s Representation Theorem) 42 Let L be any finite distributive lattice, and J L the subset of join-irreducibles. ⊆ J is a partially ordered set (not a sublattice!). The family of lower sets in J is a lattice of sets, and isomorphic to L ([Bi1], Corollary III.3.2).
Of particular interest for us is L = Lip X for some metric space X, with and + ∧ pointwise minimum and maximum respectively, and , pointwise infimum ∨ and supremum. Lip X is a distributive lattice, as the distributivity is inherited V W from (R, , ). The following proposition is a special case of Lemma 6.3 in ∧ ∨ [He] and Proposition 1.5.5 in [Wv].
Proposition 43 ...... 43 Let X be a metric+ space. Then Lip X is complete as a lattice.
Proof Let J be some arbitrary index set, and let f be in Lip X for each j J. j ∈ Obviously, [0, ] is complete as a lattice, with = and = 0. So we ∞ ∞ define pointwise ∅ ∅ V W
g(x) := fj(x), h(x) := fj(x) j J j J _∈ ^∈ and observe that g and h : X Z are Lipschitz: For arbitrary x,y X holds → ∈ h(x) f (x) f (y) + d(x,y) ≤ j ≤ j for all j J, and thus, by passing to the infimum: ∈ h(x) h(y) + d(x,y). ≤ Same for g.
Example 44 ...... 44 The space C([0, 1], [0, 1]) of real-valued continuous functions [0, 1] [0, 1] is → not a complete lattice with pointwise minimum and maximum: Choose fj(x)= 0 (1 j x), j N , the infimum is not continuous. The same example shows ∨ − · ∈ ∗ that the space K 0 LipK,0 X of all Lipschitz-functions with arbitrary Lipschitz constant is not a complete≥ lattice. S
Example 45 ...... 45 Besides its Stone representation, we want to provide another, more intuitive Basic Notions in Order Lattices 21 representation of Lip X by a lattice of sets, using its hypograph (cp. “epigraph” in [Ro])
hyp : Lip X ℘ (X [0, ]) → × ∞ f (x, r) : f(x) r . 7→ { ≤ } (im hyp, , ) obviously is isomorphic to (Lip X, , ) as a lattice; however, ∩ ∪ ∧ ∨ they are not yet isomorphic as complete lattices: Infinite unions of the closed sets in im hyp are not closed in general – we have to use the union with closure “ ¯” instead of the traditional union. (Alternatively, we could identify subsets ∪ of X [0, ] with the same closure.) × ∞ 22 Claim 45 Chapter 1
Order Lattices with Metrics
1.1 Valuation Metric Lattices
For the most part of this thesis, we will consider the supremum metric+
d (f, g) := f(x) g(x) . ∞ − x X _∈ on Lip X. But before we come to further investigate this, we present Birkhoff’s [Bi1] definition of a metric lattice and demonstrate the difference to our situa- tion.
Definition 46 ...... 46 A valuation on a lattice L is a function v : L R which satisfies the modular → law
v(f) + v(g) = v(f g) + v(f g) f, g L. ∧ ∨ ∀ ∈ A valuation v on L is called isotone [positive] if for all f, g L the relation ∈ f < g implies v(f) v(g) [v(f) < v(g)]. ≤ If L is totally ordered, then each function v : L R is a valuation. It is isotone → [positive] if and only if v is [strictly] monotonically increasing.
Example 47 ...... 47 Let L = (N∗, gcd, lcm). Then each logarithm is a positive valuation on L.
Example 48 ...... 48 Let V be any finite dimensional vector space, and L = PG(V ) its lattice of sub- vector spaces, with the intersection and the span (the projective geometry ∧ ∨ of V ). Then the dimension function is a positive valuation on L.
Historical Remark The theory of valuations has been mainly developed and popularized by John von Neumann and Garrett Birkhoff. In the early years of 24 Claim 49 the 1930s, von Neumann worked on a variation of the ergodic hypothesis, and inadvertently competed with George David Birkhoff. Only some years later, his son Garrett Birkhoff pointed von Neumann at the use of lattice theory in Hilbert spaces. He wrote about this in a note of the Bulletin of the AMS in 1958 [Bi2].
John von Neumann’s brilliant mind blazed over lattice theory like a meteor, during a brief period centering around 1935–1937. With the aim of interesting him in lattices, I had called his attention, in 1933– 1934, to the fact that the sublattice generated by three subspaces of Hilbert space (or any other vector space) contained 28 subspaces in general, to the analogy between dimension and measure, and to the characterization of projective geometries as irreducible, finite- dimensional, complemented modular lattices. As soon as the relevance of lattices to linear manifolds in Hilbert space was pointed out, he began to consider how he could use lat- tices to classify the factors of operator-algebras. One can get some impression of the initial impact of lattice concepts on his thinking about this classification problem by reading the introduction of [...], in which a systematic lattice-theoretic classification of the different possibilities was initiated. [...] However, von Neumann was not content with considering lattice theory from the point of view of such applications alone. With his keen sense for axiomatics, he quickly also made a series of funda- mental contributions to pure lattice theory.
The modular law in its earliest form (as dimension function) appears in two papers from 1936 by Glivenko and von Neumann ([Gl], [vN]). Von Neumann used it (and lattice theory in general) in his paper to define and study Continu- ous Geometry (aka. “pointless geometry”), and later applied his knowledge to found Quantum Logic in his Mathematical Foundations of Quantum Mechanics. A later survey about metric posets is [Mn].
Example 49 ...... 49 Let (X, Σ,µ) be a probability space. The σ-algebra Σ is a Boolean lattice by union and intersection. Let c R be arbitrary, then ∈ v(A) := µ(A)+ c
defines an isotone valuation on Σ with v( ) = c. The valuation v is positive if ∅ and only if there are no null sets in X other than . ∅ Proof: Let A = be a null set. Then ( A, but µ( )=0= µ(A). Conversely, 6 ∅ ∅ ∅ if A ( B Σ, and v(A)= v(B), then define C := B r A. B is a disjoint union ∈ of A and C, so by σ-additivity of µ holds µ(A) = µ(B) = µ(A)+ µ(C), hence µ(C) = 0. Valuation Metric Lattices 25
The distance function d (A, B) := v(A B) v(A B) (whose properties will v ∪ − ∩ be proved in Lemma 53) is the measure of the symmetric difference A △ B of A and B, if A △ B Σ. It relates to the Hausdorff distance just as the 1-distance ∈ of functions relates to the supremum distance.
We further exploit the connection between dv and the symmetric difference. The symmetric difference as an operation makes sense only in complemented lattices (power sets are examples for complemented lattices). But although all distributive lattices can be represented by a lattice of sets, and hence can be embedded into a complemented lattice, they need not be complemented by themself.
Example 50 ...... 50 Let L = (Z, min, max). A possible representation of L is the family
A(L) := ( ,n] Z n Z { −∞ ∩ | ∈ } of subsets of Z, equipped with union and intersection. This lattice is order- isomorphic to L. The canonical distance in Z is given by the L-valuation v(n) := n, as well as by the counting measure of the symmetric difference of sets in A(L). However, L is distributive, but not complemented, there is no proba- bility measure µ on A(L) which corresponds to v, and the symmetric difference of two sets in A(L) does not yield a set of A(L) again; these problems are all connected to each other, as they all base on the fact that L is not complemented. They still persist even if one completes L to L := (Z , min, max) and ′ ∪ {±∞} generalizes v to a valuation with infinity.
Note that the existence of symmetric differences f △ g, of complements f c, and of difference sets f r g are equivalent to each other in complete lattices of sets:
f c := 1 △ f = 1 r f with 1 := f f L _∈ f △ g := (f g)c (f g) = (f r g) (g r f) ∧ ∧ ∨ ∪ f r g := f gc = f △ (f g) ∧ ∧ In particular, complement, symmetric difference, and difference are unique. Lip X never is a complemented lattice. Hence, there is no difference defined on Lip X; however, given a valuation v and f, g Lip X, we may define a difference ∈ valuation of f and g by:
w(f, g) := v(f) v(f g). − ∧ As f g f, we conclude that for positive or isotone v, the difference valuation ∧ ≤ w always is non-negative.
Proposition 51 (Difference valuation properties) ...... 51 Let v be an isotone valuation on a distributive lattice L and w its difference valuation. Then the following holds for all f,g,h L: ∈ 26 Claim 51
1. w(f, g) + w(g, f) = v(f g) v(f g) =: d (f, g) ∨ − ∧ v 2. w(f, g) = w(f, g h) + w(f h, g) (cut law) ∨ ∧ 3. f g w(f, g) = 0 ≤ ⇒ 4. v is positive if and only if for all f, g L holds: w(f, g) = 0 f g. ∈ ⇔ ≤ 5. Let w : L L R be a map satisfying property (2), c R arbitrary, and × → ∈ let 0 L be a least element. Then v(f) := w(f, 0) + c is a valuation with ∈ w as difference valuation, and all valuations with w as difference valuation are of this form. If w(f, g) 0 for all f, g L, then v is isotone. ≥ ∈ Proof (1) Simply by the defining property of v: w(f, g)+ w(g, f) = v(f) 2v(f g)+ v(g) − ∧ = v(f g) v(f g) ∨ − ∧ (2) By the defining property of v we know: v(f h)+ v(f g) = v ((f g) (f h)) + v ((f g) (f h)) ∧ ∧ ∧ ∨ ∧ ∧ ∧ ∧ = v (f (g h)) + v(f g h) ∧ ∨ ∧ ∧ Inserting this into the definition of w yields: w(f, g) = v(f) v(f g) − ∧ = v(f) v (f (g h)) + v(f h) v(f g h) − ∧ ∨ ∧ − ∧ ∧ = w(f, g h)+ w(f h, g) ∨ ∧ We call this equation “cut law” in view of its meaning in Venn diagrams (see Figure 1.1). (3) Using (2): f = f g w(f, g)= w(f g, g)+ w(f, g g) = 2 w(f, g). ∧ ⇒ ∧ ∨ (4, “ ”) f g is equivalent to f g = f, and by positivity, equivalent to ⇒ ≤ ∧ v(f g)= v(f). ∧ (4, “ ”) Let f < g. Then f = g f, and ⇐ ∧ v(g) v(f) = v(g) v(g f) = w(g,f). − − ∧ By isotony we know v(f) v(g), assume v(f) = v(g), then w(g,f) = 0, and ≤ hence g f, contradicting f < g. ≤ (5) First of all we deduce an easy consequence of properties (2) and (3): w(f g, g) = w((f g) f, g)+ w(f g,f g) = w(f, g) ∨ ∨ ∧ ∨ ∨ With this at hand, we can easily conclude that v is a valuation and that w is its difference valuation: v(f) = w(f, 0) + c = w(f g, 0) + w(f, g)+ c ∧ = v(f g) + w(f, g) ∧ and v(f g) = w(f g, 0) + c = w(g, 0) + w(f g, g) + c ∨ ∨ ∨ = v(g) + w(f, g) v(f) + v(g) = v(f g) + v(f g) ⇒ ∨ ∧ Valuation Metric Lattices 27
f g f g f g
h h h w(f, g)= w(f, g v h) + w(f v h, g)
Figure 1.1: Visualization of the cut law of difference valuations (Proposition 51.2), using Venn diagrams. Using a representation π, the set π(f) r π(g) is cut along π(h) to give π(f) r π(g h) and π(f h) r π(g). ∨ ∧
Let w(f, g) 0, f g, then v(g) v(f)= v(g) v(f g) 0, i.e. v is isotone. ≥ ≤ − − ∧ ≥ Now let v be any valuation with w as difference valuation, then
v(f) = v(f) v(f 0) + v(0) = w(f, 0) + v(0) − ∧ obviously holds, choose c = v(0).
Proposition 51.5 shows the equivalence of the concepts of valuation and dif- ference valuation for complete lattices, so we may define the term “difference valuation” without reference to an actual valuation:
Definition 52 ...... 52 A difference valuation on a distributive lattice L is a function w : L L R × → which satisfies the cut law
w(f, g) = w(f, g h) + w(f h, g). ∨ ∧ A difference valuation w is called isotone if its values are non-negative, and positive, if w(f, g) = 0 implies f g. ≤
The following Lemma is a part of Theorem X.1 and a note in subsection X.2 of [Bi1], and can equally well be stated in terms of valuations as well as difference valuations:
Lemma 53 ...... 53 Let v be an isotone valuation on the distributive lattice L. Then
d (f, g) := v(f g) v(f g) v ∨ − ∧ defines a pseudo-metric with the following properties:
1. If there is a least element 0 L, then ∈ v(f) = v(0) + d (f, 0) for all f L, v ∈ 28 Claim 53
f g f g f g f g f g
d(f, g) = + + + h h h h h
f g f g f g
d(f, h) = + h h h
f g f g f g
d(h, g) = + h h h
Figure 1.2: Proof of the triangle inequality for valuation metric lattices, e.g. of Lipschitz function spaces with L1-metric. Note that f,g,h are functions in this case, and symbolized by sets via Stone duality.
2. dv is a metric if and only if v is positive.
We call dv a valuation (pseudo-)metric. A lattice together with a valuation metric is sometimes called a metric lattice; however, as we will deal with lat- tices with non-valuation metrics as well (particularly the supremum metric), we should better distinguish between valuation metric lattices and non-valuation metric lattices.
Proof dv(f,f) = 0, dv(f, g) = dv(g,f) and property (1) are obvious. The absorption laws tell us that f f g and f f g for all f, g L, hence ≤ ∨ ≥ ∧ ∈ f g f g and isotony of v yield d (f, g) 0. Contrary to Birkhoff, we will ∨ ≥ ∧ v ≥ use difference valuations to prove triangle inequality:
d (f, g) = w(f, g h)+ w(f h, g)+ w(g, f h)+ w(g h, f) v ∨ ∧ ∨ ∧ Due to positivity of w and property (2) in Proposition 51, we have
w(f, g h) w(f, h) ∨ ≤ w(f h, g) w(h, g) ∧ ≤ w(g,f h) w(g, h) ∨ ≤ w(g h, f) w(h, f) ∧ ≤ And thus
d (f, g) w(f, h)+ w(h, f)+ w(h, g)+ w(g, h) d (f, h)+ d (h, g). v ≤ ≤ v v
We finally show that dv is a metric if and only if v is positive. Again, we use dv(f, g) = w(f, g)+ w(g,f) to see that dv(f, g) = 0 implies w(f, g) = 0 and w(g,f) = 0. Proposition 51, property (4) applies: f g and g f, thus ≤ ≤ Valuation Metric Lattices 29
f g f g f g
h h h
v d(f, g)= d(f v h, g v h) + d(f v h, g h)
f g f g f g
h h h
v
d(f, g)= d(f v (g v h), f (g v h))
v + d(g v (f v h), g (f v h))
f g f g f g
v d(f, g)= d(f, f g) + d(f v g, g)
=d(g, f v g) + d(f v g, f)
Figure 1.3: Demonstration of the use of Venn diagrams to prove calculation rules in valuation metric lattices. The final example shows how a single Venn diagram can be interpreted in two different ways to match two different rules. f = f g = g. For the other direction, keep in mind that f g w(f, g) = 0 ∧ ≤ ⇒ always holds. So, assume that dv is a metric, and w(f, g) = 0 though f g. Then f = f g, d (f,f g) > 0, and w(f g,f)= d (f,f g) w(f, g) > 0. 6 ∧ v ∧ ∧ v ∧ − But f g f, so w(f, g) = 0, contradiction. ∧ ≤
The Lemma still holds for non-distributive lattices, just property (1) is weaker: d (f g,f h) + d (f g,f h) d(g, h) for all f,g,h L. v ∨ ∨ v ∧ ∧ ≤ ∈ The proof uses the fact that even in a non-distributive lattice, there still holds a distributive inequality. These calculations can easily be visualized by Venn diagrams. Each element f L corresponds to a set in R2, and the valuation v equals the area of this set. ∈ The distance between f and g then can be seen as the area of the symmetric difference of f and g. Now the union of the symmetric differences of f and h on the one hand, and g and h on the other hand, is a superset of the symmetric difference of f and g. Hence, d (f, g) d (f, h) + d (h, g). v ≤ v v However, this metaphor is not a full proof, as the operation of symmetric dif- ference fails for general lattices. We will return to this matter in Lemma 64 in a more general context. 30 Claim 55
g x g x g x g x
f y f y f y f y d(f v g, x v y) + d(f v g, x v y) d(f, x)+ d(g, y)
Figure 1.4: Another property of valuation metrics, visualized with a four-set Venn diagram; a precursor to Proposition 72.
Example 54 ...... 54 We continue with the special case L = (Z, min, max) from example 50. Although it was not possible to create a link between the valuation v and the counting measure µ, the difference valuation w on L provides us with a useful connection:
w(n,m) = µ(A(n) r A(m)) In cases of distributive lattices without least elements, it is feasible and beneficial to forget about v and define dv directly in terms of the difference valuation. Proposition 51 shows that this is possible without loss of information.
We want to demonstrate the use of Venn diagrams to derive calculation rules in valuation metric lattices:
Proposition 55 ...... 55 Let (L, , ) be a distributive lattice with difference valuation w and corre- ∧ ∨ sponding valuation metric d. Then for all f, g, x, y L holds: ∈ d(f g, x y) + d(f g, x y) d(f, x) + d(f, y) ∧ ∧ ∨ ∨ ≤ Proof Figure 1.4 depicts the proposition. As the symmetry of the figure suggests, we can prove the slightly stronger statement
w(f g, x y) + w(f g, x y) w(f, x) + w(f, y) (*) ∧ ∧ ∨ ∨ ≤ from which the thesis directly follows. Next, we decompose each term into smaller subterms with the help of the cut and absorption laws, to gain subterms of the form w(a . . . a , b . . . b ) in which each of f, g, x, y appear at 1 ∧ ∧ n 1 ∨ ∨ n exactly one position. These are minimal terms, which each correspond to one of the fifteen minimal areas in the Venn diagram. Figure 1.4 helps us to determine along which function we have to cut. We get:
w(f g, x y) = w(f g x, y) + w(f g y, x) + w(f g, x y) ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∨ w(f g, x y) = w(f g, x y) + w(f, g x y) + w(g, f x y) ∨ ∨ ∧ ∨ ∨ ∨ ∨ ∨ w(f, x) = w(f g y, x) + w(f y, g x) ∧ ∧ ∧ ∨ + w(f g, x y) + w(f, g x y) ∧ ∨ ∨ ∨ w(g, y) = w(f g x, y) + w(g x, f y) ∧ ∧ ∧ ∨ + w(f g, x y) + w(g, f x y) ∧ ∨ ∨ ∨ Valuation Metric Lattices 31
It is possible to recombine the remaining subterms and express the difference of left- and right-hand side as another sum of distances—we leave this as recre- ation to the reader.
Example 56 ...... 56 If L = Lip X, with X a measure space, we may apply the Lebesgue integral to gain an isotone valuation on L; as f + g = (f g) + (f g) holds pointwise, we ∧ ∨ conclude
f dµ + g dµ = (f g)dµ + (f g)dµ. ∧ ∨ Z Z Z Z If X is a Euclidean space, or a discrete space without non-trivial null sets, this valuation is positive, because any non-trivial non-negative Lipschitz function has positive Lebesgue integral. Positivity fails in cases where X contains an isolated point or continuum of measure zero. As f g = (f g) (f g) holds pointwise, the valuation metric d equals | − | ∨ − ∧ v the L1-distance defined by
d (f, g) := f g dµ. 1 | − | Z This raises the question for which metrics d on Lip X there is a valuation v, such that d = dv. From property (2) in Lemma 53 we can easily deduce the valuation v. Insertion into the definition of a valuation leads to the requirement d (f, 0) + d (g, 0) = d (f g, 0) + d (f g, 0) f, g Lip X. v v v ∨ v ∧ ∀ ∈ In the special case of functions f, g with disjoint support we have f g = f + g, ∨ so we end up with the necessary condition f g = 0 d (f, 0) + d (g, 0) = d (f + g, 0). ∧ ⇒ v v v Now it should not take us any wonder that “completely non-linear” metrics like the supremum metric or
d (f, g) := p f g p dµ p | − | sZ for p> 1 are non-valuation metrics.
Example 57 ...... 57 Yet, there are some more valuation metrics besides the L1-metric. Let κ : [0, ) [0, ) be a positive valuation (i.e., strictly monotonically increasing), ∞ → ∞ then
vµ,κ(f) := κ(f(x)) dµ(x) Z is a positive valuation. However, the author is not aware of any valuation metric on Lip X, which cannot be described in this way with suitable κ and µ. 32 Claim 60
1.2 Ultravaluation Metrics
Lemma 58 ...... 58 Let L be a distributive lattice, and let w : L L [0, ] be a map which × → ∞ satisfies
(1) w(f, g) = 0 whenever f g, and ≤ (2) w(f, g) = w(f h, g) w(f, g h) f,g,h L. ∧ ∨ ∨ ∀ ∈
We call w a difference ultravaluation, or just ultravaluation. Define dw(f, g) := w(f, g) w(g,f). Then d is a pseudo-ultrametric . d is an ultrametric if ∨ w + w + and only if w(f, g) = 0 f g holds. ⇒ ≤
Proof To get from normal valuations to ultravaluations, we just replaced all occurences of “+” by “ ”. As both operations are associative and commutative, ∨ we can transfer most proofs of valuations just by replacing “+” by “ ”: ∨ d (f, g) = w(f, g h) w(f h, g) w(g,f h) w(g h, f) w ∨ ∨ ∧ ∨ ∨ ∨ ∧ w(f, g h) w(f, h) etc. ∨ ≤ d (f, g) w(f, h) w(h, g) w(g, h) w(h, f) = d (f, h) d (h, g) ⇒ w ≤ ∨ ∨ ∨ w ∨ w
On the other hand, contrary to the valuation case, the property dv(f,f) = 0 does not follow from property (2) – we have to conclude it from (1). Assume w(f, g) = 0 f g holds. Let d (f, g) = 0. This implies w(f, g) = 0 ⇒ ≤ w and w(g,f) = 0, and hence f g, g f, and f = g. Now assume d is a ≤ ≤ w metric, f g, and w(f, g) =0. Then
w(f,f g) = w(f g,f g) w(f, g) = 0. ∧ ∧ ∧ ∨ Due to f g, we have f = f g, hence 6 ∧ 0 < d (f,f g) = w(f,f g) w(f g,f) = w(f g,f). w ∧ ∧ ∨ ∧ ∧ But f g f, contradiction. ∧ ≤
Example 59 ...... 59 Let X be any set, κ : X [0, ] arbitrary and fixed, and L a lattice of subsets → ∞ of X. For A, B L consider ∈ w(A, B) := 0 sup κ(x). ∨ x ArB ∈ w defines an ultravaluation. Choose κ to be a positive constant, then the ultrametric resulting from w will be the discrete metric on X. Ultravaluation Metrics 33
Lemma 60 ...... 60 Let X be finite, and let L be a lattice of subsets of X. Then any ultravaluation on L is of the form of Example 59.
Proof For x X and A, B X define ∈ ⊆ κ(x) := inf w(C, D) : C, D L with x C, x / D { ∈ ∈ ∈ } and w′(A, B) := 0 sup κ(y). ∨ y ArB ∈ Assume w (A, B) > w(A, B). Then there is y A r B with κ(y) w(A, B), ′ ∈ ≥ but this cannot happen, as one may choose C = A and D = B. Hence, assume w (A, B) < w(A, B). Then for all y A r B there should be C, D L with ′ ∈ ∈ y C r D and w(C, D) < w(A, B). As ∈ w(C, D) w(C A, D B), ≥ ∧ ∨ we might choose without loss of generality C A and D B, as choosing ⊆ ⊇ C A instead of C and D B instead of D further decreases w(C, D). The cut ∩ ∪ law now yields
w(A, B) = w(C D, B) w(C, D) w(A D, B C) w(A, C D). ∧ ∨ ∨ ∨ ∨ ∨ ∨ As w(C, D) < w(A, B) by assumption, we find that at least one of (C D)rB, ∩ (A D) r (B C), and A r (C D) must be non-empty. Choose y out of their ∪ ∪ ∪ ′ union and repeat the above argument for the now smaller subset. We get an infinite sequence of different elements from X, which is a contradiction because X is finite.
Example 61 ...... 61 Now consider L = Lip X, with X some metric+ space. We represent L with subsets of X [0, ], as described in Example 45, to apply it to Example 59. The × ∞ most canonical κ would be κ = π , the projection onto [0, ]. The corresponding 2 ∞ ultrametric on L is
d (f, g) = 0 sup f(x) g(x) with x X such that f(x) = g(x) . κ ∨ { ∨ ∈ 6 } We shall call this metric the “peak metric” on Lip X. Another possible choice for κ is as follows: Choose a basepoint x X and 0 ∈ κ(x, r) := dX (x,x0). Then dκ will describe the greatest distance from x0 at which f and g still differ. Finally, κ(x, r) := exp( d (x,x )) will describe − X 0 the least distance from x0 at which f and g differ. We will call the first case the “outer basepoint metric” and the second case the “inner basepoint metric”. An application of the lower basepoint metric is as follows: Given a free group F with neutral element x0, identify each normal subgroup N E F with its characteristic function on F . These are 1-Lipschitz functions in the canonical word metric of F . dκ then defines a topology on Lip F , which restricts to the Cayley topology ([dH], V.10) on the subset of normal subgroups. 34 Claim 63
1.3 Intervaluation Metrics
We now integrate the supremum metric into the context of difference valuation, but not without a sincere generalization of the concept. Similar to the case of the ultravaluation, We first recognize the possibility to replace “+” in the definition of a difference valuation by any commutative and associative binary operation. But this alone will not suffice to encompass the supremum metric, we have to weaken the main property of a difference evaluation as well:
Definition 62 ...... 62 An intervaluation on a distributive lattice (L, , ) is a map w : L [0, ] ∧ ∨ → ∞ together with a commutative and associative binary operation : [0, ] ◦w ∞ × [0, ] [0, ], such that the following properties hold: ∞ → ∞ 1. r 0 = 0 r = r ◦w ◦w 2. r t (r + s) (t + u) (r t) + (s u) ◦w ≤ ◦w ≤ ◦w ◦w 3. r s r s (follows from (1) and (2)) ∨ ≤ ◦w 4. f g w(f, g) = 0 ≤ ⇒ 5. w(f, g h) w(f h, g) w(f, g) w(f, g h) + w(f h, g) ∨ ◦w ∧ ≤ ≤ ∨ ∧ (left and right modular inequality, or cut law) for all f,g,h L and r,s,t,u [0, ]. The corresponding intervaluation metric ∈ ∈ ∞ then is defined to be
d (f, g) := w(f, g) w(g, f). w ◦w The intervaluation is positive if
w(f, g) = 0 f g. ⇒ ≤
Proposition 63 ...... 63 An intervaluation w on L and its metric dw always fulfill:
1. w(f, g) = w(f g, g) = w(f, f g) = d (f g, g) f, g L. ∨ ∧ w ∨ ∀ ∈
2. dw is a pseudo-metric+ .
3. dw is a metric+ if and only if w is positive.
Proof (1) We choose h = f or h = g in both modular inequalities:
0 w(f, g) w(f g, g) 0 + w(f, g) ◦w ≤ ∨ ≤ w(f, g) 0 w(f, g f) w(f, g) + 0 ◦w ≤ ∧ ≤ and d (f g, g) = w(f g, g) 0 = w(f, g). w ∨ ∨ ◦w Intervaluation Metrics 35
(2) From the definition we see d (f, g) 0 and d (f, f) = 0 for all f, g L. w ≥ w ∈ As is commutative, d is symmetric. ◦w w d (f, g) = w(f, g) w(g, f) w ◦w (w (f h, g) + w (f, g h)) (w (g h, f) + w (g, f h)) ≤ ∧ ∨ ◦w ∧ ∨ (w (h, g) + w (f, h)) (w (h, f) + w (g, h))) ≤ ◦w = (w (f, h) + w (h, g)) (w (h, f) + w (g, h))) ◦w (w (f, h) w (h, f)) + (w (h, g) w (g, h))) ≤ ◦w ◦w = dw(f, h) + dw(h, g)
(3, “ ”) Assume 0 = w(f, g) = w(f, f g). Then d (f,f g)=0+0=0. ⇒ ∧ w ∧ As d is a metric, we have f = f g, so f g. w ∧ ≤ (3, “ ”) d (f, g) = 0 implies w(f, g)=0 and w(g, f) = 0, hence f g f, ⇐ w ≤ ≤ and f = g.
The definition of intervaluations is chosen to generalize difference valuations and ultravaluations, while maintaining as many inequalities as possible. As those calculation laws derived from Venn-diagrams hold for difference valuations and ultravaluations likewise, which both work as the extremal cases of intervalua- tions, it is a worthwhile endeavour to figure out those laws that still hold for intervaluation metrics, which contain lots of interesting metrics for Lipschitz function spaces.
Lemma 64 ...... 64 Let L be a distributive lattice with intervaluation metric+ d. Define a polyno- mial in L as an expression of finitely many variables in L, linked by and . ∨ ∧ Define a difference term in L to be a term of the form w(x, y), where x and y are polynomials in L. Define a difference polynomial in L as a term built from finitely many difference terms, linked by addition. Let and be difference T1 T2 terms. Let f to f be the set of all variables occuring in and . Draw a 1 N T1 T2 Venn diagram with f1 to fN as basis, and let π be the correspondence between the subsets of the Venn diagram and polynomials of f1 to fN in L. Define ρ(x, y) := π(x) r π(y). Define π( ) to be the union of all ρ(x, y) for which T1 w(x, y) appears in . If π( ) π(T ), then (2N 1) . T1 T1 ⊆ 2 T1 ≤ − ·T2
Proof Call the 2N 1 minimal subsets in the Venn diagram atoms. Each atom − A is uniquely described by a non-empty subset S of f ,..., f , such that { 1 N }
A = π(f ) r π(f ) , j j fj S fj / S \∈ [∈ and hence A = ρ( S fj, Sc fj). On the other hand, each difference term w(x, y) can be uniquely chopped down into corresponding atoms of the form V W w( S fj, Sc fj) as well, by repeatedly using the cut law, up to N times (for V W 36 Claim 65 each variable once; as and distribute over each other as well as over them- ∧ ∨ selves, the absorption law reduces each polynomial to a polynomial of or ∧ ∨ only); uniqueness follows from commutativity. By Definition 62 we conclude: maximum of up to (2N 1) atoms w(x, y) sum of up to (2N 1) atoms − ≤ ≤ −
As j J aj #J maxj J aj for any finite index set J and real numbers aj, ≤ · ∈ we have∈ P sum of atoms (2N 1) w(x, y) (2N 1) (sum of atoms) ≤ − · ≤ − · Correspondingly, each difference term can be bounded from above and below T by multiples of the atoms of its constituents. If π( ) π( ), then each atom T1 ⊆ T2 of the left-hand Venn diagram occurs in the right-hand Venn diagram as well, hence the sum of atoms in is a subsum of the sum of atoms in . We then T1 T2 follow sum of -atoms sum of -atoms (2N 1) . T1 ≤ T1 ≤ T2 ≤ − ·T2
Example 65 ...... 65 There are several possible choices for the commutative and associative binary operation in Definition 62. Choosing addition leads directly to the definition ◦w of valuations. The next important choice is the maximum operation: Properties (1) and (3) are obviously fulfilled, the left side of (2) as well. (2.right) needs some short consideration: As + distributes over , the right-hand side equals ∨ (r t) + (s u) = (r + s) (r + u) (t + s) (t + u) ∨ ∨ ∨ ∨ ∨ which is greater or equal (r + s) (t + u) for all r, s, t, u [0, ]. ∨ ∈ ∞ Each norm on R2 with certain normalization properties qualifies as an || · || operation via r s := (r, s) . This accounts for the p-norms: ◦w ◦w || || r s := (r, s) := √p rp + sp ◦p p for p [1, ). Again, properties (1), (2.left) and (3) of Definition 62 are triv- ∈ ∞ ial. Property (2.right) is the triangle inequality of the p-norms (i.e. a special case of the Minkowski inequality [Wr]).
Given any metric d on L we may define w (f, g) := d(f g, g) and deduce d ∨ ◦w from d(f, g) = w (f, g) w (g, f). The operation must be commutative d ◦w d ◦w due to the symmetry of dw. From the remaining properties of Definition 62, property (4) follows directly from d(g, g) = 0, while the rest is less obvious.
Example 66 ...... 66 The standard metric on [0, ] is an intervaluation metric with + ∞ + 0 (r s) : s < w(r, s) := ∨ − ∞ r, s [0, ]. 0 : s = ∀ ∈ ∞ ∞ Intervaluation Metrics 37
However, one may freely choose to be addition or maximum. To prove the ◦w cut law for both choices, it suffices to show
0 (r s) = 0 r (s t) + 0 (r t) s) . ∨ − ∨ − ∨ ∨ ∧ − For this, we make use of a + b = (a b) + (a b) with a = r s and b = r t, ∧ ∨ ∧ ∧ then add r to both sides, rearrange and apply x (x y) = 0 (x y). − ∧ ∨ − Example 67 ...... 67 Let (X, µ) be a measure space, L = Lip X, and p (1, ) arbitrary. Define ∈ ∞ r s = (rp + sp)1/p, and ◦w
w(f, g) := p f (f g) p dµ . − ∧ sZ
As r (r s) p + s (r s) p = r s p for all r, s [0, ] (with p := ), | − ∧ | | − ∧ | | − | p ∈ ∞ ∞ ∞ the corresponding (pseudo-)metric+ is just the L -metric+
d (f, g) = p f g p dµ . p | − | sZ Properties (1)-(3) of Definition 62 follow from Example 65, (4) is trivial. The left cut law can be shown by pointwise analysis and case distinction (h g vs. ≤ h > g), the right cut law follows from Example 66 and the Minkowski inequality. dp might be a pseudo-metric+ , depending on µ.
Example 68 ...... 68 Here is a minimal example for a non-intervaluation metric: Take L = a, b, c { } with a
LC(f g) LC(f g) LC(f) LC(g) ∨ ∨ ∧ ≤ ∨ 38 Claim 71 the inverse inequality is wrong, as there is no bound to LC(f) by any combi- nation of LC(f g) and LC(f g). To see this, consider the two-point-space ∧ ∨ X = a, b of diameter l < 1, and the Lipschitz-functions f = (0, l) and { } g = (l, 0). Then LC(f) = f = 1, but LC(f g) = LC(f g) = 0 and || ||L ∧ ∨ = l in both cases. || · ||L Correspondingly, the cut law is explicitly violated by dLC, as one can see when f and g are two different constant functions, and h crosses them both.
We now concentrate on the special case of the supremum metric.
Proposition 70 ...... 70 Let Z be a distributive lattice with intervaluation metric+ d (with corresponding w and ), with r s = r s for all r, s [0, ]. Let X be an arbitrary d ◦d ◦d ∨ ∈ ∞ space, and L a complete lattice of functions f : X Z with pointwise infima → and suprema. Then
w (f, g) := wd f(x), g(x) ∞ x X _∈ defines an intervaluation metric+ on L with r s = r s for all r, s [0, ], ◦∞ ∨ ∈ ∞ which equals the supremum metric+ d . ∞ Proof The left inequality of the cut law is trivial. For the right side we have to use that a supremum of sums is less than or equal to a sum of suprema, which in turn follows from complete distributivity:
w fx, gx w fx, (g h)(x) + w (f h)(x), gx d ≤ d ∨ d ∧ x X x X _∈ _∈ w fx, (g h)(x) + w (f h)(x), gx ≤ d ∨ d ∧ x X x X _∈ _∈
Corollary 71 ...... 71 Let X be any metric+ space. The supremum metric+ d is an intervaluation ∞ metric+ on Lip X.
Proof We have seen in Proposition 43 that Lip X is a complete lattice. We easily find r s = r s and ◦d∞ ∨ w (f, g) = f(x) (f g)(x) = 0 f(x) g(x) , d∞ − ∧ ∨ − x X x X _∈ _∈ which is the intervaluation metric+ of Proposition 70 applied to Example 66.
Apart from all laws which we may derive from Venn diagrams, the supremum metric on Lip X provides yet another interesting law, which is an adaptation and generalization of Proposition 55: Complete Metric/Lattice-Irreducibility 39
Proposition 72 ...... 72 For f , g arbitrary set theoretic functions X [0, ], j J, J some arbitrary j j → ∞ ∈ index set, holds:
d fj, gj d (fj, gj) ∞ ≤ ∞ j J j J j J ^∈ ^∈ _∈
d fj, gj d (fj, gj) ∞ ≤ ∞ j J j J j J _∈ _∈ _∈
Proof For J = , both inequalities are trivial. Assume J = . As j J and ∅ 6 ∅ ∈ x X commute, it suffices to show ∈ W W d xj, yj d (xj,yj) ∞ ≤ ∞ ^ ^ _ d xj, yj d (xj,yj) ∞ ≤ ∞ _ _ _ for any x ,y [0, ]. j j ∈ ∞ First we handle infinities. First inequality: Assume there is j with x = y = . j j ∞ We can ignore all such j’s from J, unless all x and y are . In this case on j j ∞ both sides are zeros. Now assume x = = y . Then appears on the right j ∞ 6 j ∞ side and trivializes the inequality. So we can restrict to finite xj and yj. Note that x = can only happen when all x = . j j ∞ j ∞ Second inequality: Assume x = , but y is finite. Then there is an V j j ∞ j j upper bound for y but not for x . Hence the right side becomes infinite, j W j W too. Note that infinite xj or yj automatically lead to infinite j xj or j yj, respectively. W W Without restriction let x y , and let M := d(x ,y ). Let δ > 0 be j j ≥ j j j j j arbitrary. Then there is an m J with y y + δ. Furthermore we have V ∈V m ≤ j j W d(x ,y ) M, hence y x M. Altogether: m m ≤ m ≥ m − V x x y + M + δ j ≤ m ≤ j Now let δ 0. The other^ inequality works^ the same way. →
1.4 Complete Metric/Lattice-Irreducibility
Recall Definition 36 of a join-irreducible element p in a lattice L:
p = f g p = f or p = g f, g L ∨ ⇒ ∀ ∈ Let L be equipped with the discrete metric ddis. Then the above property is equivalent to the following:
d (p, f) d (p, g) d (p, f g) f, g L dis ∧ dis ≤ dis ∨ ∀ ∈ 40 Claim 75
In the same sense, p is completely irreducible if and only if
ddis p, fj ddis p, fj (fj)j J L, J = . ≤ ∀ ∈ ⊆ 6 ∅ j J j J ^∈ _∈ Definition 73 ...... 73 Let L be a complete lattice with intervaluation metric d. We call p L com- + ∈ pletely ml-irreducible if the following equivalent conditions hold:
1. (fj)j J L, with J an arbitrary non-empty index set: ∀ ∈ ⊆
d p, f d p, f j ≤ j j J j J ^∈ _∈
2. (fj)j J L, J an arbitrary non-empty index set, and R [0, ] : ∀ ∈ ⊆ ∀ ∈ ∞
d p, f R δ > 0 j J : d (p,f ) R + δ j ≤ ⇒ ∀ ∃ ∈ j ≤ j J _∈ (This results from expanding the infimum in (1).)
Let 0 L be the least element in L, then a bounded completely ml-irreducible ∈ element p L is a completely ml-irreducible element with finite distance to 0. ∈ Denote the subset of L of all [bounded] completely ml-irreducible elements with cmli(L) [bcmli(L)].
Proposition 74 ...... 74 Let L be metrically closed, then each completely ml-irreducible element is join- irreducible. It is not necessarily completely join-irreducible (see Definition 36).
Proof Use R = 0 and that a set of two elements is compact. For a counter- example to complete join-irreducibility, let L = [0, 1] with standard metric, supremum and infimum. Take f = 1 1/n, n N , then p =1 = f , n − ∈ ∗ n hence p is not completely join-irreducible. Still, it is completely ml-irreducible: W Any sequence of real numbers fn with p = fn must converge to p from below, hence d(p, f ) = 0. n W V
Proposition 75 ...... 75 Let L be a complete lattice with intervaluation metric+ d, and let L be metrically complete. Then cmli(L) is topologically closed. If 0 is the least element of L, then bcmli(L) is topologically closed, and 0 bcmli(L). ∈ Complete Metric/Lattice-Irreducibility 41
Proof Let (p ) cmli(L), n N be some sequence of completely ml-irre- n ⊆ ∈ ∗ ducible elements converging to p L, and (fj)j J any non-empty family in L. ∈ ∈ Then for any n N holds ∈ ∗
d p, f d p , f d(p, p ) j ≥ n j − n _ d(p _, f ) d(p, p ) ≥ n j − n ^ d(p, f ) d(p, p ) d(p, p ) ≥ j − n − n ^ d (p, f ) 2 d(p, p ), ≥ j − n 0 ^ → | {z } i.e. the element p is completely ml-irreducible. Any component of L is topologically closed, in particular the component of 0, hence the intersection with cmli(L) is closed as well. Furthermore, we have for any k J = : ∈ 6 ∅
d 0, fj = wd fj, 0 idempotency _ = w f_ f , 0 cut along f d k ∨ j k w (f , 0) _ w f ,f ≥ d k ◦w d j k w (f , 0) = d(0_, f ) ≥ d k k and thus d(0,f ) d(0, f ) d 0, f . j ≤ k ≤ j V W
Definition 76 ...... 76 Let L be a lattice with metric d, R 0 arbitrary. We define an R-base of + ≥ L to be a subset B L such that for any f L there is (bj)j J B, J an ⊆ ∈ ∈ ⊆ arbitrary non-empty index set, such that d(f, b ) R.A base simply j J j ≤ is a 0-base. Given a least element 0 L, a bounded∈ R-base is an R-base of L ∈ W which is a subset of the component of 0.
Example 77 ...... 77 Let L = [0, ] with standard metric d (modulus of the difference). Then one ∞ + easily calculates cmli(L) = bcmli(L) = [0, ). Furthermore, any metrically ∞ dense subset B of [0, ) is a bounded base, in particular, we have sequences ∞ (b ) B with b = , although d(b , ) = , and thus b in d. j ⊆ j ∞ j ∞ ∞ j 6→ ∞ W Proposition 78 ...... 78 Consider an R-base B of a complete lattice L with intervaluation metric+ d, R 0. Then for each δ > 0, cmli(L) is in the (R + δ)-ball around B. In ≥ particular, if R = 0, cmli(L) lies in the metrical closure of B. 42 Claim 79
Proof Let p cmli(L) be arbitrary. As B is an R-base, there are b B, ∈ j ∈ j J = , such that ∈ 6 ∅
d p, b R. j ≤ j J _∈ From Definition 73 we infer that there is a sequence (c ) B, k K J k ⊆ ∈ ⊆ whose distances to p converge to R. If R = 0, the sequence (cj) metrically converges to p.
Example 79 ...... 79 It is easy to see that, if B is a base, and b B not a join-irreducible element, ∈ then B r b is a base as well (if b = f g, f and g are joins of elements of { } ∨ B, and as f,g < b, b is not part of these joins). Using the Lemma of Zorn, it is possible to deduce that the subset of all join-irreducible elements constitutes a base for any sufficiently nice lattice.
Unfortunately, this is not the case with ml-irreducible elements: Let L′ be the completely distributive complete lattice [0, 3] [0, 2] with componentwise supre- × mum and infimum, and with supremum metric. Then consider the sublattice L L formed by the five elements ⊆ ′ L := (0, 0), (1, 0), (0, 1), (1, 1), (2, 2) . { } We find cmli(L) = (0, 0), (1, 0), (0, 1) , as (1, 1) = (1, 0) (0, 1). p = { } ∨ (2, 2) is join-irreducible in this lattice, but not ml-irreducible: Take f1 = (1, 0), f2 = (0, 1), then d(p, fj) = 2, but d(p, fj) = 1. Nevertheless, (2, 2) must be part of any 0-base of L. V W Chapter 2
Rough Isometries of Lipschitz Function Spaces
2.1 Smoothening of Lipschitz Functions
Definition 80 ...... 80 Let X and Z be metric spaces, and ǫ, K 0 fixed. We define the Z-valued + ≥ K-Lipschitz and (K,ǫ)-Lipschitz functions on X by:
Lip (X, Z) := f : X Z : x,y X : d(fx,fy) K d(x,y) K { → ∀ ∈ ≤ } Lip (X, Z) := f : X Z : x,y X : d(fx,fy) K d(x,y)+ ǫ K,ǫ { → ∀ ∈ ≤ } Later on we will restrict to Z = R and Z = [0, ]. ∞ Given two metric spaces X, Y and a (λ, ǫ)-quasi-isometry η : X Y , we can + → push η to a map between the Lipschitz function spaces
η∗ : LipK,δ(Y, Z) LipK λ, K ǫ+δ(X, Z) → · · f f η. 7→ ◦ Using the supremum metric (aka. L∞-metric) d on Lipschitz functions, we ∞ easily see that η∗ is a (2Kǫ + 2δ)-isometric embedding of LipK,δ(Y, Z) into LipK λ, K ǫ+δ(X, Z): Obviously, we have · · d(fηx,gηx) d (f, g) x X. ≤ ∞ ∀ ∈ On the other hand, for each 0 D < d (f, g) we may find y Y with ≤ ∞ ∈ d(fy,gy) D, and x X with d(ηx,y) ǫ. Hence, ≥ ∈ ≤ d(fηx,gηx) d(fy,gy) d(fηx,fy) d(gηx,gy) ≥ − − D 2 Kǫ 2 δ. ≥ − − Taking the supremum over x, we conclude d (η∗f, η∗g) d (f, g) 2Kǫ 2δ. ∞ ≥ ∞ − −
Question Is η∗ even a quasi-isometry? 44 Claim 81
We may split this problem into two special cases: The rough case with λ = 1. And the bilipschitz case ǫ = 0, for which we can give an easy counterexample: Let X be R2 with Euclidean metric, and let Y be R2 with supremum metric. The identity is a 2-bilipschitz bijection. Let Z = R and f : X Z, (x ,x ) → 1 2 7→ x2 + x2, i.e. f(x)= x , and f Lip (X, Z). In particular, for points x with 1 2 | |X ∈ 1 x = x we find f(x)= √2 x . But for g Lip (Y, Z) must hold p1 2 · 1 ∈ 1 g(x) g(0) sup x ,x , | − | ≤ { 1 2} such that with x , f and g have infinite distance from each other. One 1 → ∞ would argue that in this case, we have to choose g Lip (Y, Z). But then, ∈ √2 we may choose points x = (x , 0) to see that the function (x ,x ) √2 x 1 1 2 7→ · | |Y has infinite distance to any f Lip (X, Z). To solve this problem, one has to ∈ 1 generalize Lipschitz functions to allow for varying Lipschitz constants – however, our main focus lies on the rough isometry case λ = 1. As
Lip (X, Z) Lip (X, Z) K ⊆ K,ǫ for any ǫ 0, we may reformulate our question in the following way: ≥ Question
How L∞-dense is the subset of all K-Lipschitz functions in the metric+ space of (K, ǫ)-Lipschitz functions?
We will next give an answer to this problem in the special case Z = R. A similar result for continuous functions is given by Petersen in [P], section 4.
Lemma 81 ...... 81 Let X be a separable metric space, K,ǫ > 0 and let f : X R be a function + → with
fx fy K d(x,y) + ǫ. | − | ≤ · Then there is a K-Lipschitz function g : X R with → d (f, g) 2 ǫ. ∞ ≤
Proof Let f : X R bea(K,ǫ)-Lipschitz function, K,ǫ 0. Let aj j N∗ → ≥ { } ∈ ⊆ X be a countable dense subset. Define g(a1) := f(a1). Now define inductively and prove by induction the following:
• j 1 − I := g(a ) K d(a , a ), g(a )+ K d(a , a ) j k − · j k k · j k k=1 \ Smoothening of Lipschitz Functions 45
• f(a ) : f(a ) I j j ∈ j g(a ) := min I : f(a ) min I j j j ≤ j max I : f(a ) max I j j ≥ j (a): g a1,..., aj is K-Lipschitz. • |{ } (b): f(a ) g(a ) ǫ. • | j − j |≤ We see that (a) and (b) are trivially fulfilled for j = 1. At first we have to show that Ij = : As g a1,..., aj−1 is K-Lipschitz, we know for all m,n g(a ) g(a ) K d(a , a ) k < j | j − k |≤ · j k ∀ and together with the Lipschitz property of g on a1,..., aj 1 we see that g { − } is also K-Lipschitz on a ,..., a . { 1 j} Finally, we show f(a ) g(a ) ǫ. If f(a ) I , we have nothing to show. | j − j | ≤ j ∈ j Assume the second case: f(a ) min I and g(a ) = min I . Let n be such that j ≤ j j j f(a ) g(a ) = min I = g(a ) K d(a , a ) j ≤ j j n − · n j By definition of Ij we have g(a ) K d(a , a ) g(a ) K d(a , a ) m < j n − · n j ≥ m − · m j ∀ for some n, and we choose n to be the least possible index with this property. Furthermore, we have f(a ) f(a ) K d(a , a )+ ǫ | j − n | ≤ · j n f(a ) f(a ) K d(a , a ) ǫ. ⇒ j ≥ n − · j n − We use this to calculate 0 g(a ) f(a ) = g(a ) K d(a , a ) f(a ) ≤ j − j n − · n j − j g(a ) K d(a , a ) f(a )+ K d(a , a )+ ǫ ≤ n − · n j − n · n j g(a ) f(a )+ ǫ. ≤ n − n Now assume g(an) >f(an). Then there is some m g(a ) = g(a ) K d(a , a ) n m − · n m g(a ) K d(a , a ) = g(a ) K d(a , a ) K d(a , a ) ⇒ n − · n j m − · n m − · n j g(a ) K d(a , a ) ≤ m − · m j 46 Claim 82 which contradicts the minimality of n. Hence we have g(a ) f(a ) and n ≤ n g(a ) f(a ) ǫ. | j − j |≤ The third case (f(aj) > g(aj) = max Ij) works the same: g(a ) = g(a )+ K d(a , a ) n smallest possible j n · n j 0 f(a ) g(a ) = f(a ) g(a ) K d(a , a ) ≤ j − j j − n − · n j f(a ) f(a )+ K d(a , a )+ ǫ | j ≤ n · n j f(a ) g(a )+ ǫ ≤ n − n Assume f(a ) > g(a ), then m